Skip to content
Koralis AIBook a free Scoping Review

Kovac · in beta · for IT, security and AI leads

AI for your whole workforce, one team at a time.

Say yes to one AI project on confidential data.

The business keeps asking for AI on the files you have kept away from AI: rates, client names, contracts. You decide which columns the AI must never see, with whoever owns the data, and that is your policy.

Book a free Scoping Review← Other teams

With both founders, no data. Pilots from £5,000 + VAT + your AI model's usage. Half at the start, and half when we send your payback report.

The third request this month for AI on the client files arrives, this time from the head of bids. The files carry client names and day rates under NDA. So the only answer you could defend has been no.

An illustration, not a customer's story.

One request, from your files to the result

IT, security and AI leads: one request, start to finishYour material: Your data classification (a document, by end Oct 2026) and The business's dataset, as files (a record set). Fields you mark: Client names, Day rates, Commercial terms. Kovac checks requests to your data against your policy. From late October 2026, the AI model you choose drafts a bid answer from the team's past bids, working through that check. On every AI call the fields you mark reach it masked: Client names arrives as [CLIENT_1]. Unmarked fields, and the text inside Word and PDF documents, arrive as they are. The result: A first-draft bid answer to edit. Separately, today: A writer not cleared for day rates asks for day rates, a marked field. Kovac refuses the request and records the refusal. People search that record today; from late October 2026, the AI model can also draft from it. Solid lines work today; dashed lines are dated.Your materialYour data classificationby end Oct 2026The business's dataset, as filesFields you markClient namesDay ratesCommercial termsKovac checks against your policyMarked fields masked on every AI callfrom late Oct 2026The AI model you chooseDrafta bid answer from the team's past bids[CLIENT_1]in place of client namesUnmarked fields and Word or PDF text asthey areThe resultA first-draft bid answer to editClearance decides what a person may read.A writer not cleared for dayratesasks for day ratesRefusedKovac checks against your policy, todayRefused, and recordedA writer not cleared for day ratesasked for day ratesPeople search it today.From late Oct 2026, the AI model candraft from it.
  • Works today
  • The AI model's part, from late Oct 2026; Raw files and previews, by end Oct 2026
The detail: the example, what it works on, and what stays governed

The bids team asks for AI on the bid library

The head of bids asks IT to approve an AI model on an export of past bids, with the CVs left out. You and the head of bids agree the policy: client names and day rates are marked. The model receives the bid register with those columns as placeholders, and drafts from it and from past bids in Word, whose text reaches it as written. A bid writer not cleared for day rates who asks for that column is refused, and the refusal is recorded.

Fields you mark sensitive reach the AI model masked with placeholders; fields you don't mark reach it as they are. The same value always gets the same placeholder, so masked data can still be linked to a person and remains personal data.

On records loaded into Kovac, someone not cleared for a column you have marked who asks for it is refused, and the refusal is recorded.

Clearance decides what a person may read. It never unmasks a field for the AI model: a marked field reaches the model masked on every call, whoever asks.

What it works on

Document
Your AI and acceptable-use policies
Document
Your data classification
Document
The business's request for an AI project
Records
The export the business wants AI on

Columns this team would mark

  • Client names
  • Contract and day rates
  • Commercial terms held in columns
  • Any column your classification calls confidential

What the AI does, and what your team does

From late October 2026, the AI model you choose, working through Kovac's check. Each card says who does it. Your team can search Kovac's record of refusals today; once pilots run, the model can draft from it too. Kovac does not draft or search; its check decides what the model may see.

Draft

The AI model does this

A monthly note of the requests Kovac refused on the bids dataset, drafted from its record of refusals

More it could draft
  • A policy in plain terms for one dataset, from your data classification, for you and the business owner to agree
  • The approval note for the project: what the AI may read, which fields it receives masked, and who is refused
  • A note of the requests Kovac refused, by dataset and column, from its record of refusals

Find

Your team does this

Which requests were refused, by whom and for which column, searched in Kovac's record by your own team

More the AI model could find
  • The columns in a proposed export that your data classification calls confidential
  • The clauses in your acceptable-use policy that a proposed AI project would rely on

Check

The AI model does this

A proposed project's export against your policy, column by column, before you say yes

More it could check
  • A proposed AI project's export against your policy, column by column, before you say yes
  • The policy against your classification, listing confidential columns it finds unmarked, for your team to check

Questions your team will ask

How does this sit with the AI already in our office suite?

Beside it, not in front of it. Kovac governs only the data you load into it. What your office-suite AI can see is still set by your own labels and permissions. See the limit: what Kovac does not sit in front of

Who installs it, and who patches and monitors it?

For a pilot, we set Kovac up on your own machines. A hosted model receives unmarked fields and document text as they are; your own keeps them on your network. Patching, monitoring and your MSP's view are not settled yet.

What happens if Kovac's check itself fails?

Not every failure blocks a request yet; the limits below say what happens today and from when. Read that line before you approve anything. See the limit: when a check fails

How do people sign in, and what does it connect to?

Sign-in is by OpenID Connect (OIDC) only today. No SAML single sign-on or SCIM provisioning. No date yet. For a pilot, the data arrives as files. PostgreSQL and MySQL connect today, for schema discovery only. See the limit: when live reads arrive

Signing it off?

One page on what Koralis AI holds, what it does not, and a date for each gap where there is one. Read it, or send it to whoever reviews suppliers.

Email the brief to a colleagueRead the one-page IT & legal brief →

What to bring to the Scoping Review

No data: descriptions are enough.

  • The AI request you keep turning down, and who is asking
  • Your data classification and AI acceptable-use policy, as they stand
  • Who in your team or your MSP would look after Kovac's machine

What it does not do yet, and what a pilot is

  • Masking works on the fields you mark: the columns of a spreadsheet or record set. It does not reach into the text of a Word or PDF document.
  • Three failure cases still let a request through when they should stop it: a deploy whose policy call fails, a read of a record whose clearance lookup errors, and a plan that cannot be signed, which runs unsigned rather than halting. Every route blocks a request when the policy check cannot run, by the end of October 2026.
  • Kovac does not sit in front of Microsoft Copilot, Gemini for Google Workspace or web chat assistants.
The other eight limits

What the check does today

  • Today the check covers three of seven routes to data; opening raw files and previews is checked by the end of October 2026.
  • Kovac's record holds the decisions on the routes it checks: refusals today; permitted requests too by the end of October 2026.
  • The record is searchable, but any signed-in user can add an entry in anyone's name, and entries are not signed or linked together when written. By the end of October 2026, only Kovac itself can write to the record; by then, every entry is signed and linked to the one before it as it is written, permitted requests as well as refusals.

What a pilot is

  • The first step is one team: a four-week pilot on your own machines, with one dataset, supplied as files, one workflow, and the AI model you choose working through Kovac's check from late October 2026.
  • Reading straight from your database tables comes in December 2026.

What we do not hold yet

  • Kovac is in beta. Koralis AI holds no certification today.
  • Professional indemnity insurance is not held today.
  • A pilot on personal data waits for a data processing agreement, which has no date yet.

Bring the AI project you keep saying no to.

Bring the refusal it would have to pass, and describe the data. If a pilot follows, we set Kovac up locally, on your own machines. Both founders, no data, a written pack within 48 hours.

The terms

  • The Scoping Review is free. It signs you to nothing.
  • Four-week paid pilots, with Kovac on your own machines, from late October 2026.
  • Quick-Start, for mid-sized businesses: £5,000 + VAT, fixed.
  • Proof Pilot, for regulated and larger organisations: £7,500 + VAT, fixed.
  • The pilot's job is to give you a payback figure, measured on your own work.
  • Half at the start, and half when we send your payback report. By invoice.
  • You pay for your AI model's usage, either to its provider directly or passed through by us.
  • Stop at the day-five go/no-go and you pay only the first half.
  • If you subscribe within 90 days of the pilot ending, the pilot fee you've paid is credited against your first year's subscription.
  • Your contract is with Koralis AI Limited (company number 16558134).
  • Kovac is not yet in production anywhere.
Book a free Scoping Review

Rather write first? enquiries@koralis.ai reaches both founders. Or message Dylan or Marko on LinkedIn.